JiraThings it finds
Repeated work often leaves a clear pattern. Browse twelve anonymized examples, with counts as recorded.
The same prep sheet asked for, plus 14 transcripts pasted in by hand, in 26 days. The calendar already linked both.
Star it, open Notion, retype the title, paste the link. About 48 a day, every day, for nearly three years.
Look up the account, post, paste the link back, chase. 2.5 days each, 694 more tickets behind them.
"Bug or by design?", "who owns this?" In 90 days, answered from four people's heads. Nothing written down.
9 of 9 call transcripts already sat in Drive. 4 got a write-up, typed out afterwards. The overlap between one meeting and the next surfaced 12 times in two weeks.
School event, dentist, flight. Read the mail, open the calendar, type it again. 2,096 times before anyone called it a process.
Structural validation of the changelog page, run manually in 7 sessions and 8 PRs in 30 days. Real, but plumbing on top of what already works.
Meeting ends, transcript saved, summarized, sorted by topic into a folder. One person did it 4,409 times over two and a half years.
New sign-up lands. Look up the company, the size, the role; paste it into the CRM; ping sales. Eight tabs, every time.
Recording to notes by hand, action items never reaching a tracker, a Slack thread relayed into Notion. Five jobs across three teams' spaces.
Skim, tag, file the useful ones. 5,935 issues in two years. A summary and a tag would have done the same job.
In a call, at lunch, out. Set by hand from the same calendar that already knows. 1,993 times.
It also finds work that already runs and skips work that is not worth building. Names are removed; counts match the source records.
Start with the work you know
Start with one source and a clear time window. Review the findings, then approve each new source before the skill reads it. Every proposal links to records you can check.
Learn from workflow patterns
A taxonomy of 907,353 n8n workflow rows helped define task patterns. Feedback refined the instructions; model weights stayed fixed. We tested the procedure on separate synthetic company data.
Built from workflow patterns. Tested on separate synthetic data.
n8n's workflow taxonomy: 907,353 workflow rows across 100,282 instances, turned into task definitions.
Feedback on task, citations, change and count reshaped the instruction text. Model weights stayed fixed.
57 unseen synthetic companies, two reader models, strict checks. Empty reports score zero.
These results compare the released procedure with its starting version on synthetic records. They do not measure the full interactive flow or time saved by real teams.
We tested a later version on 116 more synthetic companies. It did not pass the release rule, so version 1.0.0 keeps the confirmed procedure. A separate comparison with a short prompt did not show a clear gain in proposal quality. Read the whitepaper for the full method and its limits.
Good to know
Install the skill
npx skills@latest add n8n-io/automation-mining-skill --skill automation-miningThen start with either prompt:
Works with Claude Code, Codex, Cursor, Pi, and other hosts that support Agent Skills.